amt-augmentor 1.0.9

Audio + MIDI paired augmentation toolkit for Automatic Music Transcription (AMT)

AMT-Augmentor

Python Data Augmentation Toolkit for Automatic Music Transcription (AMT)

Developed by Bots for Music, maintained by Lars Monstad

Note: Formerly known as amt-augpy1.0 (a.k.a. amt-augpy). Starting with v1.0.9 the package is published on PyPI as amt-augmentor and the import path is amt_augmentor.

A Python toolkit for augmenting Automatic Music Transcription (AMT) datasets through various audio transformations while maintaining synchronization between audio and MIDI files. The dataset follows the same format as MAESTRO v3.0.0, which is commonly used for Automatic Music Transcription (AMT) tasks.

The toolkit expects a folder containing paired audio and MIDI files with matching names. The audio file and MIDI file must be ground truth data, as this toolkit is only for augmenting existing datasets - a common technique in Machine Learning.

dataset/
├── song1.wav        # Audio file
├── song1.mid        # Ground truth annotated midi file

Features

Audio Transformations

• Time Stretching: Tempo modification while maintaining pitch
• Pitch Shifting: Transposition while preserving timing
• Reverb & Filtering: Room acoustics and frequency filtering effects
• Gain & Chorus: Depth and richness enhancement
• Noise Augmentation: Controlled noise addition for robustness training
• Pause Manipulation: Detection and modification of musical pauses

Processing & Dataset Handling

• Audio Standardization: Conversion to 44.1kHz WAV format
• Parallel Processing: Multi-core processing for faster augmentation
• Configuration System: YAML-based parameter customization
• Dataset Validation: Automatic validation of train/test/validation splits
• MAESTRO Compatibility: Dataset format compatible with MAESTRO v3.0.0

Installation

You can install AMT-Augmentor either via pip or by cloning the repository:

Using pip

pip install amt-augmentor

From source

git clone https://github.com/LarsMonstad/amt-augmentor.git
cd amt-augmentor
pip install -e .

For development, install with additional development dependencies:

pip install -e ".[dev]"

Dependencies

• librosa
• soundfile
• numpy
• pedalboard
• pretty_midi
• pyyaml
• tqdm

Usage

Basic Usage

amt-augmentor /path/to/dataset/directory
# Or running directly
python -m amt_augmentor.main /path/to/dataset/directory

This will process all compatible audio files in the directory and their corresponding MIDI files. The script automatically selects random parameters within predefined ranges for each augmentation type.

Advanced Usage

# Use a custom configuration file
amt-augmentor /path/to/dataset/directory --config my_config.yaml

# Specify an output directory
amt-augmentor /path/to/dataset/directory --output-directory /path/to/output

# Generate a default configuration file
amt-augmentor --generate-config my_config.yaml

# Disable specific effects
amt-augmentor /path/to/dataset/directory --disable-effect timestretch --disable-effect chorus

# Parallel processing with 8 workers
amt-augmentor /path/to/dataset/directory --num-workers 8

# Custom train/test/validation split
amt-augmentor /path/to/dataset/directory --train-ratio 0.8 --test-ratio 0.1 --validation-ratio 0.1

# Force specific songs to test set (prevents augmentation)
amt-augmentor /path/to/dataset/directory --custom-test-songs "song1,song3,song5"

# Dry run to preview what will be processed
amt-augmentor /path/to/dataset/directory --dry-run

# Verbose output for debugging
amt-augmentor /path/to/dataset/directory --verbose

# Check for valid MIDI-WAV pairs before processing
amt-augmentor /path/to/dataset/directory --check-pairs

# List available effects
amt-augmentor --list-effects

# Check version
amt-augmentor --version

Help and options

amt-augmentor --help

Configuration

All augmentation parameters can be customized using a YAML configuration file. See config.sample.yaml for a complete example with documentation.

Sample Configuration

# Time stretching configuration
time_stretch:
  enabled: true
  variations: 3
  min_factor: 0.6
  max_factor: 1.6

# Pitch shifting configuration
pitch_shift:
  enabled: true
  variations: 3
  min_semitones: -5
  max_semitones: 5

# Noise addition configuration
add_noise:
  enabled: true
  variations: 3
  min_intensity: 1.1
  max_intensity: 2.0
  randomized: true

# Processing configuration
processing:
  num_workers: 4
  output_dir: null

File Format Support

Audio

• Input: WAV, FLAC, MP3, M4A, AIFF
• Output: WAV (44.1kHz)

Annotations

• MIDI (.mid)

Output Structure

For each input file pair (audio + MIDI), the toolkit generates multiple augmented versions with the following naming convention:

original_name_augmented_effect_parameter_randomsuffix.extension

The _augmented_ identifier ensures all augmented files are properly recognized and handled during dataset creation.

Example:

piano_augmented_timestretch_1.2_abc123.wav
piano_augmented_timestretch_1.2_abc123.mid
piano_augmented_noise_1.5_def456.wav
piano_augmented_noise_1.5_def456.mid

Dataset Creation & Validation

The dataset follows the same format as MAESTRO v3.0.0. Songs assigned to test or validation splits will have their augmented versions excluded to prevent data leakage.

Creating the Dataset CSV

# Create dataset with default split ratios (70% train, 15% test, 15% validation)
amt-augmentor /path/to/directory

# Create dataset with custom split ratios
amt-augmentor /path/to/directory --train-ratio 0.8 --test-ratio 0.1 --validation-ratio 0.1

# Force specific songs to test set (they won't be augmented)
amt-augmentor /path/to/directory --custom-test-songs "song1,song3,song5"

Validating the Dataset Split

Dataset split validation is automatically performed after CSV creation to ensure:

• Augmented songs are not included in test/validation splits
• No cross-split contamination occurs
• Original and augmented songs are properly distributed

CSV Format

The generated CSV follows the MAESTRO format with the following columns:

• canonical_composer
• canonical_title
• split
• year
• midi_filename
• audio_filename
• duration

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

For development:

1. Install development dependencies: pip install -e ".[dev]"
2. Run tests: pytest tests/
3. Check typing: mypy amt_augmentor
4. Format code: black amt_augmentor

Contributors

• Lars Monstad (@LarsMonstad) – Original author and maintainer
• @monoamine11231 – Noise augmentation, custom test songs feature, and various improvements

Contact

For questions or collaboration:

• Email: [e-mail]
• Organization: https://botsformusic.com
• GitHub: https://github.com/LarsMonstad/amt-augmentor

License

MIT License - see LICENSE file for details.

Citation

If you use this toolkit in your research, please cite:

@software{amt_augmentor,
  author       = {Lars Monstad and contributors},
  title        = {AMT-Augmentor: Audio + MIDI augmentation toolkit for AMT datasets},
  version      = {1.0.9},
  year         = {2025},
  publisher    = {Bots for Music},
  url          = {https://github.com/LarsMonstad/amt-augmentor}
}